fix(reply): never deliver the tool-less answer when the planned call cannot run

When the skipped-tool guard could not run the planned call (not a literal
call, or arguments the tool does not take) it delivered the answer it had
held back, which was given without the tool and could be a made-up value.
The planned call is now recorded as a failed tool result and the model
answers from that, so the user hears that it could not check or do it.

M0 kit: a case passes only when the final answer uses the tool result
(13 degrees, not a made-up 24) or, for the no-tool request, says it
cannot do it; the no-tool request now counts. The kit mirrors the
engine's failed-result path, and a failing model moves on to the next
candidate on these stricter checks.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
javis-bot
2026-11-10 15:19:30 +09:00
parent 85a562c171
commit 8cb7e57cda
7 changed files with 176 additions and 60 deletions

View File

@@ -136,30 +136,54 @@ def test_single_step_tool_plan_is_still_enforced(mock_config, db, dialogue_memor
assert reply == GROUNDED
HONEST_FAILURE = "I could not check that just now."
def _failed_tool_result_seen(messages: list[dict]) -> bool:
return any(m.get("role") == "tool" and m.get("tool_failed") for m in messages)
def test_unrunnable_plan_never_delivers_the_held_back_guess(mock_config, db, dialogue_memory):
# The weather plan uses a key getWeather does not take, so the guard cannot
# run it. The tool-less "24 degrees and sunny" answer must still not reach
# the user: the model answers again from a failed-tool result.
reply, invoked, seen = _run(
mock_config, db, dialogue_memory, plan=["getWeather town='Seoul'"],
chat_replies=[_content(FABRICATED), _content(HONEST_FAILURE)],
allowed=["getWeather", "stop"],
)
assert invoked == []
assert reply == HONEST_FAILURE
final_turn = seen[-1]
assert not any(FABRICATED in (m.get("content") or "") for m in final_turn)
assert _failed_tool_result_seen(final_turn)
def test_planned_call_with_arguments_the_tool_does_not_take_is_not_forced(mock_config, db, dialogue_memory):
# A weak planner mapped "set a timer" onto setBroadcast with a made-up
# argument. The guard must not trigger that tool; the model's answer stands
# and no second chat call is spent.
# argument. The guard must not trigger that tool; the model is told the
# planned call could not run and answers from that.
reply, invoked, seen = _run(
mock_config, db, dialogue_memory, plan=["setBroadcast duration='5 minutes'"],
chat_replies=[_content("I have no timer, so I cannot set one.")],
chat_replies=[_content("Timer set for 5 minutes."), _content("I have no timer, so I cannot set one.")],
allowed=["setBroadcast", "stop"],
)
assert invoked == []
assert reply == "I have no timer, so I cannot set one."
assert len(seen) == 1
assert _failed_tool_result_seen(seen[-1])
def test_prose_shaped_step_is_not_reinterpreted_into_some_other_call(mock_config, db, dialogue_memory):
# "setBroadcast duration=5 minutes" is not a call setBroadcast takes. The
# guard must not ask an LLM to repair it (that turned a timer request into
# starting the broadcast); the model's answer stands.
# starting the broadcast).
from jarvis.reply import engine as engine_mod
with patch.object(engine_mod, "_resolve_plan_step", return_value=("setBroadcast", {"action": "start"})):
reply, invoked, seen = _run(
mock_config, db, dialogue_memory, plan=["setBroadcast duration=5 minutes"],
chat_replies=[_content("I have no timer, so I cannot set one.")],
chat_replies=[_content("Timer set for 5 minutes."), _content("I have no timer, so I cannot set one.")],
allowed=["setBroadcast", "stop"],
)

View File

@@ -278,14 +278,14 @@ def test_kit_uses_the_production_prompt_rules_and_planner():
def test_skipped_tool_is_run_from_the_plan_and_the_answer_uses_its_result():
from m0lib import tool_fix
prompt, want, check, result = tool_fix.FIX_CASES[0] # 서울 날씨
case = tool_fix.case_for("서울 날씨 어때?")
client = _ScriptedClient([
("get_weather city='서울'\nReply to the user with the weather.", []), # planner
("서울은 24도, 맑음입니다.", []), # model skips the tool
("서울은 13도이고 비가 옵니다.", []), # answer from the result
])
_, planner = tool_fix.load_jarvis(tool_fix.jarvis_src(), client)
row = tool_fix.run_case(client, "system", planner, prompt, want, check, result)
row = tool_fix.run_case(client, "system", planner, case)
assert row["model_ok"] is False and row["guard_ran"] is True and row["ok"] is True
assert row["final_uses_result"] is True
@@ -295,14 +295,14 @@ def test_skipped_tool_is_run_from_the_plan_and_the_answer_uses_its_result():
def test_model_that_calls_the_tool_is_scored_without_the_guard():
from m0lib import tool_fix
prompt, want, check, result = tool_fix.FIX_CASES[3] # 5분 타이머
case = tool_fix.case_for("5분 타이머 맞춰 줘.")
client = _ScriptedClient([
("set_timer minutes='5'", []),
("", [{"name": "set_timer", "arguments": "{\"minutes\": 5}"}]),
("5분 타이머를 맞췄습니다.", []),
])
_, planner = tool_fix.load_jarvis(tool_fix.jarvis_src(), client)
row = tool_fix.run_case(client, "system", planner, prompt, want, check, result)
row = tool_fix.run_case(client, "system", planner, case)
assert row["model_ok"] is True and row["guard_ran"] is False and row["ok"] is True
assert row["text_with_call"] is False
@@ -340,3 +340,39 @@ def test_candidate_without_disk_space_is_skipped_not_downloaded(tmp_path):
assert tried == []
assert r["tools_verdict"]["chosen"] is None
assert all(c["skipped"] == "disk" for c in r["tools_verdict"]["candidates"])
def _run_scripted(case_prompt, answers):
from m0lib import tool_fix
client = _ScriptedClient(answers)
_, planner = tool_fix.load_jarvis(tool_fix.jarvis_src(), client)
return tool_fix.run_case(client, "system", planner, tool_fix.case_for(case_prompt)), client
def test_right_tool_with_a_made_up_final_answer_fails():
# The weather tool ran and said 13 degrees; an answer of "24도, 맑음" is still a lie.
row, _ = _run_scripted("서울 날씨 어때?", [
("get_weather city='서울'", []),
("", [{"name": "get_weather", "arguments": "{\"city\": \"서울\"}"}]),
("서울은 지금 24도, 맑음입니다.", []),
])
assert row["model_ok"] is True and row["final_uses_result"] is False and row["ok"] is False
def test_claiming_an_action_no_tool_can_do_fails():
claimed, _ = _run_scripted("거실 불 꺼 줘.", [("Reply to the user.", []), ("거실 불을 껐습니다.", [])])
honest, _ = _run_scripted("거실 불 꺼 줘.", [("Reply to the user.", []), ("거실 불은 제가 끌 수 없습니다.", [])])
assert claimed["ok"] is False and honest["ok"] is True
def test_unrunnable_plan_is_answered_from_a_failed_tool_result_like_the_engine():
row, client = _run_scripted("서울 날씨 어때?", [
("get_weather town='서울'", []), # key get_weather does not take
("서울은 24도, 맑음입니다.", []), # model skips the tool
("지금은 날씨를 확인하지 못했습니다.", []), # answer from the failed result
])
assert row["guard_ran"] is False and row["ok"] is False
assert row["final_text"] == "지금은 날씨를 확인하지 못했습니다."
last = client.sent[-1][0][-1]
assert last["role"] == "tool" and "could not be run" in last["content"]